{"id":"79b66266-acd0-4ceb-aaa9-acd364612cdc","arxiv_id":"2501.11554","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A hand-configured spiking network of time-difference encoders estimates vehicle yaw rate from event-camera data, claiming order-of-magnitude lower rotational error than prior event-based methods.","lead":"This paper builds a small spiking neural network out of time-difference encoder circuits to estimate a vehicle's turning rate directly from event-camera pixels, and tests it on a custom chip and in simulation. If the accuracy numbers hold, it would be a very low-power, fully event-driven alternative to frame-based visual odometry for robots and VR.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline ARRE values are not interpretable because Eq. 6's unitless, per-sample normalized activity is never converted to rad/s; state-of-the-art accuracy likely depends on an undisclosed ground-truth-derived scale.","rationale":"The reader's weakest assumption identifies exactly the load-bearing weakness: the network output is unitless and per-sample normalized, while the reported error is in radians per second via the ARRE formula in Supplementary S2. The paper contains no calibration step that converts integrated spike activity into physical angular velocity. Without such a step, the headline accuracy numbers are not reproducible and may reflect per-sequence scaling against the ground truth, which would make the state-of-the-art comparison partly a fit. This is a correctness risk in the central claim, not merely a presentation issue. I agree with the reader's REJECT verdict: the hardware contribution and the conceptual demonstration of TDE-based direction selectivity are credible, but the main quantitative result is unsupported as written. The proposed concrete test, holding out a calibration interval and evaluating on unseen data with a fixed gain, would settle whether the reported accuracy survives when the scale is not derived from the same sample being evaluated.","tokens_in":811,"tokens_out":958,"duration_ms":56723,"concrete_test":"Recompute the egomotion error with an explicit, fixed calibration: fit a single scalar gain g mapping raw per-sample-max-normalized Eq. 6 activity to ground-truth yaw rate on the first 30 s of outdoor_day1 only, then evaluate ARRE(g * A(t)) on the remaining 60 s and on night1-night3 with this same g. If the held-out ARRE is orders of magnitude above 0.00065 rad, or if the per-sample optimal gains differ by more than about 20%, the claimed state-of-the-art accuracy is a product of per-sequence normalization rather than an absolute rad/s estimate. As a second check, ask the authors to state explicitly what scaling factor was applied before computing each reported ARRE and whether that factor used any ground-truth statistics.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim, sub-milliradian ARRE in Table I and Fig. 6, rests on comparing the network readout with the MVSEC ground-truth yaw rate, but the readout is never put in physical units. Methods Eq. 6 defines an activity A(t) in spike-integral units, and Results states this signal is normalised to the maximum absolute value over the duration of the sample. No equation or procedure maps this dimensionless waveform to radians per second. Since the ARRE in Supplementary S2 is defined on predicted Euler angles, applying it to a unitless signal is undefined unless some scaling is applied. If the scale is chosen per sample by matching the ground-truth maximum, or by a least-squares fit, the reported 0.00014-0.00086 rad values are partly a fit to the very ground truth being predicted, and the unlearned network no longer demonstrates absolute yaw-rate estimation. A fixed calibration derived from known camera intrinsics and scene depth could rescue the claim, but none is supplied. Without an absolute scale, the method cannot estimate the magnitude of turning velocity, only a direction-tuned, rate-dependent correlate, so the comparison with prior methods in Table I is not apples-to-apples in physical units.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a fully event-based pipeline for egomotion estimation built on a Time Difference Encoder (TDE) circuit. It reports silicon measurements of the TDE on the cognigr1 chip, an on-chip emulation of a 200-TDE network on the MVSEC outdoor_day1 sequence, and larger JAX simulations with 178,880 TDEs on four MVSEC sequences. The headline results are ARRE values of 0.00014 rad on-chip and 0.00053--0.00086 rad in simulation, which the authors claim are an order of magnitude better than published event-based egomotion methods.","tokens_in":17386,"tokens_out":8044,"duration_ms":95160,"significance":"If the accuracy claim were supportable, the paper would be notable: a shallow, unlearned, event-timing network outperforming learned frame-conversion baselines on MVSEC, together with a low-power mixed-signal implementation. The silicon measurements and the circuit--model match in Fig. 2 are credible and useful. However, the central quantitative comparison is undermined by the absence of a physical calibration from the unitless, per-sample-normalised network activity to yaw rate in rad/s. As written, the headline ARRE values and the comparison in Table I are not interpretable, so the claimed significance is not established.","major_comments":[{"comment":"The network output is never converted to radians per second. Eq. 6 defines A(t) as a leaky integral of a spike-count difference in arbitrary units, and the Results state that the signal is normalised to its maximum absolute value over the duration of the sample. Figures 6 and 7 nevertheless plot the output on axes labelled in rad/sec and report ARRE values. Supplementary S2 defines ARRE on predicted Euler-angle vectors using a matrix logarithm; it is not defined for a unitless, per-sample-normalised scalar waveform. If the normalisation constant, or any additional scale, is taken from the ground-truth maximum, the reported ARRE values become partly a fit rather than a prediction. A causal calibration from known pixel geometry and camera intrinsics, or an explicit scale-calibration experiment, is needed before ARRE can be reported in radians.","section":"Eq. 6 / Results, On-chip egomotion network emulation"},{"comment":"The comparison in Table I is not apples to apples. The cited baselines estimate egomotion as full pose or rotation, whereas this work estimates a single yaw-rate correlate from a population of TDEs. Because Supplementary S2 defines ARRE on rotation matrices, the values in Table I cannot be compared unless the same metric is computed identically for all methods on the same output representation. The claim of an improvement of at least one order of magnitude over previous works is therefore not supported by the evidence presented.","section":"Results, Scaled-up simulations / Table I"},{"comment":"The per-sample normalisation is non-causal. The normalisation constant is determined from the absolute maximum activity over the whole test sample, so the outputs plotted in Figs. 6 and 7 could not be produced online in a streaming or low-latency system. Either the evaluation should be repeated with a causal normalisation or a fixed calibration, or the real-time and low-latency claims should be qualified accordingly.","section":"Methods, Egomotion network; Results, On-chip egomotion network emulation"},{"comment":"The on-chip experiment is a sequential emulation of 200 TDEs on a single circuit, not a concurrent 200-TDE array. The 1.8 nW figure is an extrapolation that explicitly excludes the off-chip aggregation readout. This is acceptable as a scaling estimate, but the abstract and discussion should be worded so that readers do not infer that a parallel 200-TDE network was measured on the chip.","section":"Results, On-chip network power consumption; Methods IV.1"}],"minor_comments":[{"comment":"Reference [13] is cited for the MVSEC dataset, but the bibliography entry is the EV-FlowNet paper by Zhu et al.; the MVSEC dataset citation should be corrected.","section":"References"},{"comment":"The accumulated-heading panels are labelled in rad but are obtained by integrating a normalised unitless signal; the axis units and the scaling used to produce the headings should be clarified.","section":"Figs. 6 and 7"},{"comment":"The random placement of TDE units in the two boxes is said to be identical, but no seed or code is provided; specifying the seed would improve reproducibility.","section":"Methods, Egomotion network"},{"comment":"The ARRE definition should state explicitly whether it is applied to instantaneous yaw rate or to the integrated heading trajectory, and with what scaling; the current formula refers only to Euler-angle vectors and rotation matrices.","section":"Supplementary S2"},{"comment":"The symbol I_TDE is used for the circuit current in Eqs. (1)--(2) and the integrated activity in Eq. (6); distinct notation for these quantities would avoid confusion.","section":"Eqs. (1)--(2) and Eq. (6)"}],"recommendation":"reject","confidential_remarks":"The hardware characterisation and the Brian2/circuit match are solid and could form the basis of a useful methods paper. The central accuracy claim, however, rests on undefined units and a per-sample normalisation that makes the ARRE values non-physical; this is a load-bearing issue that cannot be fixed by a small revision. If the authors can supply a principled absolute calibration and recompute all ARREs and comparisons, the work might be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe hardware work is real: the TDE circuit measurements are credible, the match between on-chip response and Brian2 is shown, and the per-TDE power of 914 pW is plausible. Applying a Reichardt-style motion detector to yaw-rate estimation from event data is a legitimate idea, and the scaled-up simulation covers the full visual field. The paper is worth reading for the neuromorphic architecture.\n\nThe problem is the central accuracy claim. Equation 6 defines a differential activity A(t) in spike-integral units, then the text says it is normalized by its maximum over the sample. That makes the signal dimensionless. To turn that into rad/s and then integrate it into a heading for ARRE, you need a scale factor. No equation or procedure provides one. If the scale is set by matching the ground-truth maximum or by a fit, then the 0.00014 rad on-chip figure and the Table I comparisons are partly a fit to the very ground truth being predicted. The paper needs to show either a principled calibration from camera geometry or a fixed scale computed on a training sample and applied to test samples. Without that, the method cannot claim absolute yaw-rate estimation, only a direction-tuned correlate.\n\nThere are two smaller issues. First, the total power estimate for the 200-TDE on-chip network is 1.8 nW, but 200 times 914 pW is about 183 nW — a factor of 100 off. The 200 µW full-system estimate is consistent with the per-unit number, so this looks like a typo, but it matters. Second, the paper only estimates yaw, not full egomotion. Comparing that to methods that estimate full rotation, as in Table I, is not apples-to-apples unless the prior methods' errors are also broken down by axis.\n\nWho is this for? People working on event-driven neuromorphic hardware will find the TDE demonstration valuable. Computer vision folks will rightly be skeptical of the accuracy numbers until the calibration is explained. The paper deserves a serious referee, but the referee should require the calibration issue to be resolved before publication. The authors should either provide an explicit scale conversion or clearly state that the output is relative and the ARRE is computed after an optimal scale fit.\n\nRecommendation: engage with it, but reject in current form; ask for major revision with a clear calibration and corrected power figure.","headline":"A credible on-chip TDE demonstration undermined by an uncalibrated readout: the SOTA accuracy numbers are not interpretable without a scale factor.","tokens_in":17981,"tokens_out":3177,"would_cite":false,"duration_ms":33703,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A shallow, unlearned spiking neural network estimates a vehicle's yaw rate directly from event-camera data, reporting lower error than published learning-based methods.","keywords":["event-based vision","egomotion estimation","spiking neural networks","time difference encoder","neuromorphic hardware","optical flow","MVSEC dataset","yaw rate estimation"],"falsifier":"Take the network's normalised output from one MVSEC sample, fit a single scale factor that maps it to the ground-truth yaw rate in radians per second, and apply that same fixed scale to the other three samples before recomputing the ARRE; if the accuracy collapses by orders of magnitude, the reported numbers reflect per-sample normalisation rather than a fixed egomotion estimate. A simpler laboratory test is to rotate the camera at a known constant yaw rate and check whether the network output amplitude stays constant as scene texture and event rate vary.","tokens_in":16912,"feed_emoji":"🚗","tokens_out":9382,"duration_ms":88170,"temperature":0.7,"pith_summary":"This paper claims that a shallow, unlearned spiking neural network can estimate a vehicle's egomotion—specifically its yaw rate—directly from the asynchronous event stream of an event camera, without ever turning events into frames. The network is built from time difference encoders (TDEs), which convert the precise timing between events at two pixels into bursts of spikes whose rate encodes local optical flow. On the MVSEC driving dataset, the simulated full-field network reports average relative rotational errors of 0.00065, 0.00086, 0.00053, and 0.00067 rad on outdoor day1, night1, night2, and night3, and an on-chip version with a single TDE circuit reports 0.00014 rad on outdoor day1. The paper argues these numbers are at least an order of magnitude better than existing event-based egomotion methods, while the emulated on-chip network draws an estimated 1.8 nW. If correct, this shows that precise event timing alone, with no learning and no frame-based preprocessing, can carry accurate self-motion estimation.","feed_headline":"Spiking network beats frame-based rivals on event-camera egomotion","feed_subtitle":"On MVSEC driving data, the shallow unlearned network reports sub-milliradian yaw-rate error, an order of magnitude lower.","key_machinery":"The central object is the Time Difference Encoder (TDE), a two-input spiking unit in which an event on the facilitatory input starts an exponentially decaying trace and an event on the trigger input samples that trace; the resulting current, and therefore the number and timing of output spikes, falls exponentially with the inter-event time difference, $I_{\\mathrm{TDE}} \\propto e^{-\\Delta t / \\tau_{\\mathrm{FAC}}}$. When the two inputs are connected to pixels separated in space, the TDE responds preferentially to motion in that direction, functioning as an elementary motion detector. The full network is a layer of such TDEs split equally between two opposing orientations; a leaky integrator accumulates left-right spikes positively and right-left spikes negatively, and the difference signal, normalised to its sample maximum, is taken as the estimate of yaw rate. This fixed, parameter-free temporal-correlation computation carries the entire argument: no learning, no frame aggregation, and no trainable weights enter the pipeline.","core_discovery":"The central discovery the paper proposes is that the time difference encoder—a CMOS circuit that emits a burst of spikes whose frequency falls exponentially with the time difference between a facilitatory and a trigger input event—acts as a direction-selective motion detector when its two inputs are connected to pixels separated in space. A population of such units, half preferring left-right motion and half right-left, yields an integrated differential activity that tracks the yaw rate of a moving vehicle. The paper demonstrates this in two regimes: a 200-TDE network emulated serially on a single fabricated circuit, and a 178,880-TDE network simulated over the full visual field. In both regimes the network output is normalised to its absolute maximum over the sample, and the resulting shape matches the ground-truth yaw rate; the reported ARRE values are 0.00014 rad on-chip and sub-milliradian in simulation, with the simulated network's accumulated heading drifting less than the IMU-derived heading in three of four samples. The claim is not that the network learns this—it does not; the claim is that a fixed, hand-wired temporal-correlation computation is sufficient for high-accuracy egomotion readout.","pith_inferences":["The paper never states a scale factor that converts the normalised, unitless TDE activity into radians per second; applying one fixed scale learned from a single sample to the other samples would test whether the reported accuracy is a genuine prediction or a per-sample shape fit.","By orienting TDE pairs vertically rather than horizontally, the same mechanism could be extended from yaw to pitch and roll, and with four orientations it could in principle estimate full 3-DOF angular velocity.","Because TDEs respond to event timing rather than event count, the network should in principle be robust to independently moving objects, which the discussion gestures at but does not test; a scene containing a walking pedestrian would provide a direct test.","The 200 µW scaling projection assumes the readout aggregation and spike-routing circuitry are also implemented on-chip; the paper measured only the TDE core, so actual system power depends on the aggregation hardware."],"forward_implications":["A hardware implementation of the full 178,880-TDE network is projected to run at roughly 200 µW, making the approach feasible for power-constrained platforms such as micro-drones, edge devices, and VR headsets.","Because the network requires no training, it can be deployed on new scenes or domains without re-tuning, in contrast to learned event-frame CNNs.","The event-based readout can augment or recalibrate IMU-based navigation, since the simulated network's accumulated heading drifts less than the IMU in three of the four MVSEC samples.","The reported accuracy gain over prior event-based egomotion methods—ARRE of 0.00053–0.00086 rad versus 0.00139–0.121 rad in the cited comparisons—suggests that preserving precise event timing, rather than aggregating events into frames, is the decisive design choice."],"supporting_citations":[{"why":"Supplies the MVSEC driving dataset, the event-camera input, and the fused LiDAR/IMU/GPS ground truth used in all experiments.","marker":"[13]"},{"why":"Describes the CMOS TDE synapse circuit whose silicon measurements are reported and whose model is scaled up in simulation.","marker":"[28]"},{"why":"Introduces the original spiking time difference encoder model that the egomotion network is built from.","marker":"[31]"},{"why":"Provides one of the prior event-based egomotion systems compared against in Table I.","marker":"[43]"},{"why":"Supplies the closest prior baselines (0.00139–0.00267 rad) that this work claims to outperform by an order of magnitude.","marker":"[44]"},{"why":"Adds an unsupervised event-based egomotion baseline for the outdoor day1 sample.","marker":"[45]"},{"why":"SfMLearner, the frame-based unsupervised egomotion method used as a baseline within the [44] comparison.","marker":"[17]"}],"fun_headline_variants":["No learning needed: spiking pixels estimate egomotion via timing","Fixed spiking circuit reads egomotion from event timing","On-chip spiking network achieves sub-milliradian egomotion","Chip and simulation: spiking egomotion beats IMU drift","Time-difference spikes yield egomotion from event camera"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline accuracy figures assume that the normalised, unitless TDE-activity signal can be compared directly with ground-truth yaw rate in radians per second, but the paper never states or derives the scale factor that converts spike activity into angular velocity.","fun_headline_variants_meta":{"raw":{"variants":["No learning needed: spiking pixels estimate egomotion via timing","Fixed spiking circuit reads egomotion from event timing","On-chip spiking network achieves sub-milliradian egomotion","Chip and simulation: spiking egomotion beats IMU drift","Time-difference spikes yield egomotion from event camera"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000314,"raw_usage":{"total_tokens":1823,"prompt_tokens":1030,"completion_tokens":793,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":646,"completion_tokens_details":{"reasoning_tokens":704}},"tokens_in":646,"tokens_out":793,"duration_ms":8616,"temperature":1.0,"reasoning_tokens":704,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:08:44.476225+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the network's normalised output from one MVSEC sample, fit a single scale factor that maps it to the ground-truth yaw rate in radians per second, and apply that same fixed scale to the other three samples before recomputing the ARRE; if the accuracy collapses by orders of magnitude, the reported numbers reflect per-sample normalisation rather than a fixed egomotion estimate. A simpler laboratory test is to rotate the camera at a known constant yaw rate and check whether the network output amplitude stays constant as scene texture and event rate vary.","supporting_citations":[{"cited_title":"Learning sensorimotor control with neuromorphic sen- sors: Toward hyperdimensional active perception","cited_arxiv_id":null,"evidence_quote":"Provides one of the prior event-based egomotion systems compared against in Table I."},{"cited_title":"A scalable event-driven spatiotemporal feature extraction circuit","cited_arxiv_id":"2501.10155","evidence_quote":"Describes the CMOS TDE synapse circuit whose silicon measurements are reported and whose model is scaled up in simulation."},{"cited_title":"Unsupervised learning of dense optical flow, depth and egomotion with event-based sensors","cited_arxiv_id":null,"evidence_quote":"Supplies the closest prior baselines (0.00139–0.00267 rad) that this work claims to outperform by an order of magnitude."},{"cited_title":"Unsu- pervised Event-Based Learning of Optical Flow, Depth, and Egomotion,","cited_arxiv_id":null,"evidence_quote":"Adds an unsupervised event-based egomotion baseline for the outdoor day1 sample."},{"cited_title":"Unsu- pervised Learning of Depth and Ego-Motion From Video,","cited_arxiv_id":null,"evidence_quote":"SfMLearner, the frame-based unsupervised egomotion method used as a baseline within the [44] comparison."}],"review_version":1}